IntelliCon: Confidence-Based Approach for Fine-Grained Vulnerability Analysis in Smart Contracts

Ethereum smart contracts are programs that execute transactions on a distributed ledger platform without intermediaries. However, they are prone to various types of vulnerabilities that can affect their security and functionality. In this paper, we present IntelliCon, a novel framework that leverages a pre-trained identifier-aware encoder-decoder CodeT5 model and confident learning to detect seven types of vulnerabilities in Ethereum smart contracts. Confident learning is a technique that improves dataset quality by identifying and correcting noisy labels, particularly in the presence of multiple annotators with varying levels of accuracy. We fine-tune CodeT5 on a dataset of 27,426 smart contracts annotated by multiple tools and pruned by confident learning to ensure that the model learns genuine vulnerability features rather than tool-specific features. Furthermore, we utilize abstract syntax tree (AST) analysis to extract code gadgets with sliding windows to locate the function that may contain code vulnerabilities. We evaluate the effectiveness of our framework in vulnerability detection with F1-score. Our results indicate that IntelliCon achieves high Micro-F1 (0.9591) and Macro-F1 (0.9293), outperforming existing methods. Moreover, our framework demonstrates its ability to handle imbalanced data, noisy labels, and complex code structures. IntelliCon contributes to improving the security and reliability of smart contracts, providing insights for future research on code generation tasks.